Introduction

Recent progress in generative artificial intelligence (gen-AI) has enabled the generation of photo-realistic and artistically-inspiring photos at a single click, catering to millions of users online. To explore how people use gen-AI models such as DALLE and StableDiffusion, it is critical to understand the themes, contents, and variations present in the AI-generated photos. In this work, we introduce TWIGMA (TWItter Generative-ai images with Metadata Annotations), a comprehensive dataset encompassing 800,000 gen-AI images collected from Jan 2021 to March 2023 on Twitter, with associated metadata (e.g., tweet text, creation date, number of likes). Through a comparative analysis of TWIGMA with natural images and human artwork, we find that gen-AI images possess distinctive characteristics and exhibit, on average, lower variability when compared to their non-gen-AI counterparts. Additionally, we find that the similarity between a gen-AI image and human images (i) is correlated with the number of likes; and (ii) can be used to identify human images that served as inspiration for the gen-AI creations. Finally, we observe a longitudinal shift in the themes of AI-generated images on Twitter, with users increasingly sharing artistically sophisticated content such as intricate human portraits, whereas their interest in simple subjects such as natural scenes and animals has decreased. Our analyses and findings underscore the significance of TWIGMA as a unique data resource for studying AI-generated images.

Dataset

You can find a more interactive introduction and exploration of our dataset at https://huggingface.co/spaces/yiquntchen/TWIGMA. Note that in accordance with the privacy and control policy of Twitter, no raw content from Twitter is included in this dataset and users could and need to retrieve the original Twitter content used for analysis using the Twitter id. Therefore, we include the following fields in TWIGMA:

• id: Twitter id uniquely identifying each tweet used in this dataset and our analysis;
• image_name: Media id (see details at the Twitter page) used to uniquely identify each photo. Leveraging this field is necessary since a tweet can contain multiple images;
• created_at: Time of creation corresponding to the Twitter id;
• like_count: Number of likes collected from official Twitter API (snapshot: the week of May 29th). Note that some likes are not available because the corresponding tweets have been deleted since we first downloaded the photos;
• quote_count: Same as like_count, but for quotes;
• reply_count: Same as like_count, but for replies;
• all_captions: This is the BLIP-generated (Li et al. 2022) captions for the corresponding image;
• label_10_cluster: This is the assigned k-means cluster (k=10 so this number varies from 1 to 10);
• possibly_sensitive: Binary variable indicating whether the media content has been marked as sensitive content by Twitter;
• nsfw_score: The predicted NSFW score from a pre-trained CLIP-based NSFW detector (ranges from 0 to 1; closer to 1 means more likely to be NSFW);
• UMAP_dim_1: The first dimension for a two-dimensional UMAP projection of the CLIP-ViT-L-14 embeddings of the images in TWIGMA.
• UMAP_dim_2: The second dimension for a two-dimensional UMAP projection of the CLIP-ViT-L-14 embeddings of the images in TWIGMA.

Link to the dataset hosted on zenodo:

Safety Note and Disclaimer:

It is important to note that a substantial amount of images in this dataset have been classified as NSFW (not-safe-for-work) by both Twitter and a CLIP-based NSFW model. This includes content that is violent, pornographic, or contains nudity. We have chosen not to exclude these images from the dataset in order to understand the content and themes without filtering. However, we have included two fields in the final TWIGMA dataset, namely "possibly_sensitive" and "nsfw_score," which can be used to filter out these images.

Please exercise caution and discretion when using this dataset. The development of this dataset has been done in compliance with Twitter's policy on data usage and sharing. The use of this dataset is solely at your own risk and should be in accordance with applicable laws, regulations, and ethical considerations. If you intend to review the original Twitter post, we recommend accessing the source page directly on Twitter and closely adhering to the official Twitter developer policy, available at https://developer.twitter.com/en/developer-terms/policy.